Estimation and Statistical Inference for Generalized Multilayer Latent Space Model

Technical Summary: Estimation and Statistical Inference for Generalized Multilayer Latent Space Model

Problem Statement
Multilayer networks, which capture multiple types of relationships among entities across different layers, are ubiquitous in fields ranging from social sciences to biology. While significant progress has been made in community detection and estimation for multilayer networks (e.g., via multilayer stochastic block models and latent space models), the theoretical framework for statistical inference remains underdeveloped. Specifically, existing inferential theory largely relies on linear models (e.g., the COSIE model) or focuses on undirected networks. There is a lack of rigorous methods for:

1. Estimating latent positions and connection matrices in directed multilayer networks with nonlinear link functions (handling binary, count, and continuous edge types).
2. Establishing the asymptotic distributions of these estimators to enable uncertainty quantification (e.g., confidence regions) and hypothesis testing (e.g., testing if two layers share the same structure).

The core challenge lies in the fact that the network structure is modeled as a tensor with a Tucker low-rank decomposition involving nonlinear link functions. This formulation renders existing techniques for linear tensor models inapplicable and introduces significant technical difficulties in deriving asymptotic normality, particularly due to identifiability issues and the non-convexity of large-scale tensor optimization.

Methodology
The authors propose a Generalized Multilayer Latent Space Model and a novel Unfolding and Fusion estimation procedure.

1. Model Formulation:

   • The model represents a multilayer directed network with $n$ nodes and $T$ layers.
   • Each node $i$ is assigned two latent positions: $\theta_i$ (sending behavior) and $\phi_i$ (receiving behavior).
   • Each layer $t$ has a specific connection matrix $\Lambda_t$ governing interactions.
   • The expected interaction $x_{ijt}$ is modeled as $x_{ijt} = \theta_i^\top \Lambda_t \phi_j + \beta_{it} + \alpha_{jt}$, where $\beta_{it}$ and $\alpha_{jt}$ account for out-degree and in-degree heterogeneity.
   • The observed edge $y_{ijt}$ follows a distribution $g_{ijt}(\cdot | x_{ijt})$ from the exponential family, allowing for binary, Poisson, or Gaussian edges via nonlinear link functions.
   • The model is expressed as a tensor $X$ admitting a Tucker decomposition: $X = [S; \Theta, \Phi, I_T] + \text{degree terms}$, where $S$ is the core tensor containing the layer-specific connection matrices.

2. Estimation (Unfolding and Fusion):

   • To avoid computationally intensive and non-convex large-scale tensor optimization, the authors propose unfolding the adjacency tensor along the first and second modes to form matrices $Z_1$ and $Z_2$.
   • Step 1: Estimate the low-rank factors $(U_m, V_m)$ for $m=1, 2$ by solving constrained maximum likelihood problems on the unfolded matrices. Constraints include orthogonality and boundedness to ensure identifiability.
   • Step 2: Apply a two-sided centering operator to remove the effects of degree heterogeneity parameters.
   • Step 3: Extract the latent position estimators ($\hat{\Theta}, \hat{\Phi}$) from the singular vectors of the centered unfolded matrices.
   • Step 4: Estimate the layer-specific connection matrices ($\hat{\Lambda}_t$) via a fusion strategy. This involves combining the right singular vectors from the first unfolding and the left singular vectors from the second unfolding through matrix multiplication, effectively reconstructing the core tensor without direct tensor optimization.

3. Identifiability:

   • The paper establishes conditions (Assumption 1) under which the parameters are identifiable up to sign flipping. This involves orthogonality constraints on latent positions and specific spectral properties of the core tensor and degree heterogeneity factors.

Key Theoretical Results
The paper establishes rigorous asymptotic properties for the proposed estimators:

1. Consistency: The estimators for latent positions ($\hat{\Theta}, \hat{\Phi}$) and connection matrices achieve uniform convergence rates. Specifically, the Frobenius norm error converges at a rate of $O_p(1/T + T^{2\delta}/n^{1-4\delta})$, and the uniform error rate is $O_p(\log n \cdot (1/\sqrt{nT} + T^\delta/n^{1-2\delta}))$.
2. Asymptotic Normality of Latent Positions: The authors derive the asymptotic normality for individual rows of the estimated latent positions ($\hat{\Theta}_i, \hat{\Phi}_i$). This is achieved by analyzing the Hessian matrix of the Lagrangian regularized log-likelihood to handle the identifiability constraints. The estimators are shown to be "oracle" estimators, achieving the Cramér-Rao lower bound.
3. Asymptotic Normality of Connection Matrices: A novel debiased inference framework is developed for the connection matrices. The authors show that the estimator $\hat{\Lambda}_t$ has a bias arising from second-order error terms in the singular vector expansions. They derive explicit forms for these bias terms and construct a debiased estimator that satisfies a central limit theorem.
4. Inference Applications: These results enable:
   • Construction of confidence regions for latent positions.
   • Hypothesis testing for structural differences between network layers (e.g., testing if $\Lambda_t = \Lambda_{t'}$).

Significance and Claims
The paper claims to be one of the first to develop inferential theory for nonlinear multilayer network models. Its primary contributions are:

• Flexibility: The model accommodates various edge types (binary, count, continuous) and directed interactions, generalizing existing models like the multilayer latent space model and COSIE.
• Computational Efficiency: The Unfolding and Fusion method circumvents the need for direct large-scale tensor optimization, leveraging established results from low-rank matrix estimation and factor analysis.
• Statistical Inference: By establishing asymptotic normality for both latent positions and connection matrices, the work provides the theoretical foundation for uncertainty quantification and hypothesis testing in multilayer networks, a gap that previously existed for nonlinear settings.
• Validation: The proposed methods are validated through extensive simulation studies and applied to real-world data, demonstrating practical utility.

The authors maintain a modest tone regarding the scope, noting that while they address the core challenges of estimation and inference for this specific class of models, the results are specific to the proposed framework and its assumptions (e.g., the factor structure for degree heterogeneity).