On weak convergence of Gaussian conditional distributions

Basic Information

• Paper ID: 2510.12412
• Title: On weak convergence of Gaussian conditional distributions
• Authors: Sarah Lumpp, Mathias Drton
• Classification: math.ST stat.TH
• Publication Date: October 14, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.12412

Abstract

Weak convergence of joint distributions does not generally imply convergence of conditional distributions. In particular, when joint Gaussian distributions converge to a singular Gaussian limit, the conditional distributions need not converge. From an algebraic perspective, this occurs because the Schur complement is not a continuous function of matrix elements at singular covariance matrices. The results in this paper elucidate special conditions under which Gaussian conditional distributions still converge, and illustrate how this enables reasoning about conditional independence in a new class of graphical models.

Research Background and Motivation

Core Problem

The core problem addressed in this research is: when a sequence of Gaussian distributions $X^{(m)} \sim N(0, Σ^{(m)})$ converges weakly to a singular Gaussian distribution $X \sim N(0, Σ)$, whether the conditional distributions $N(0, Σ_{R|S}^{(m)})$ converge to $N(0, Σ_{R|S})$.

Problem Significance

1. Theoretical Importance: The convergence of conditional distributions is a fundamental problem in probability theory, crucial for understanding the limiting behavior of distributions
2. Applied Value: In graphical models, the analysis of conditional independence relies on properties of conditional distributions, particularly in novel graphical models defined by continuous Lyapunov equations

Limitations of Existing Methods

1. Discontinuity in General Cases: The pseudoinverse operation $(·)^+$ is discontinuous at singular matrices, causing the conditional covariance matrix $Σ_{R|S} = Σ_{R,R} - Σ_{R,S}(Σ_{S,S})^+Σ_{S,R}$ to potentially diverge
2. Restrictions in Special Cases: Known results primarily apply to matrix sequences of constant rank, while practical applications frequently encounter full-rank matrices converging to singular matrices

Research Motivation

Inspired by Boege et al. (2025) on conditional independence, the authors consider the more challenging case where each $Σ^{(m)}$ is full-rank while the limiting covariance matrix $Σ$ is singular.

Core Contributions

1. Establishing Convergence Conditions: Proposes sufficient conditions for convergence of Gaussian conditional distributions, even in singular limit cases
2. Generalization of Matrix Determinant Lemma: Develops an extended version of the matrix determinant lemma for analyzing asymptotic behavior of singular matrices
3. Graphical Model Applications: Demonstrates how to apply theoretical results to conditional independence analysis in graphical models defined by continuous Lyapunov equations
4. Special Case for Toeplitz Matrices: Provides intuitive alternative proofs for special Toeplitz matrices

Methodology Details

Problem Formulation

Given a sequence of Gaussian random vectors $X^{(m)} \sim N(0, Σ^{(m)})$ converging weakly to $X \sim N(0, Σ)$, where $Σ^{(m)}$ is invertible but $Σ$ is singular. For disjoint subsets $S, R ⊆ [p]$, determine whether the conditional covariance matrices $Σ_{R|S}^{(m)}$ converge to $Σ_{R|S}$.

Core Theoretical Results

Main Theorem (Theorem 1.2)

Let $Σ^{(m)}$ be a sequence of invertible symmetric $p×p$ matrices with singular limit $Σ$, satisfying the expansion: $Σ^{(m)} = Σ + \frac{1}{m}Σ^{(1,∞)} + o\left(\frac{1}{m}\right)$

For $S ⊂ [p]$ with $k = |S| < p$, $R = [p] \setminus S$, if matrices $Σ$ and $Σ^{(1,∞)}$ satisfy:

(i) $f(U, Σ_{S,S}^{(1,∞)}) ≠ 0$, where $U$ comes from the symmetric rank decomposition $Σ_{S,S} = UBU^T$
(ii) $\text{im}(Σ_{S,R}) ⊆ \text{im}(Σ_{S,S})$

then the conditional covariance matrices $Σ_{R|S}^{(m)}$ converge to $Σ_{R|S}$.

Key Function Definition

The function $f: \mathbb{R}^{k×r} × \mathbb{R}^{k×k} → \mathbb{R}$ is defined as: $f(U,G) = (-1)^{s([k-r])} \sum_{\substack{α⊆[k] \\ |α|=k-r}} (-1)^{s(α)} \det\begin{pmatrix} G_{[k],α} & U \end{pmatrix} \det(U_{[k]\setminus α,[r]})$

Technical Innovations

1. Generalization of Matrix Determinant Lemma

Lemma 3.1: For matrices $U, V ∈ \mathbb{R}^{k×r}$, $G ∈ \mathbb{R}^{k×k}$, $B ∈ \mathbb{R}^{r×r}$, where $r < k$: $\det\left(UBV^T + \frac{1}{m}G\right) = \frac{1}{m^{k-r}} \cdot \det(B) \cdot f_{\text{asy}}(U, V, G) + o\left(\frac{1}{m^{k-r}}\right)$

2. Application of Cramer's Rule

By expressing each element of the conditional covariance matrix as: $v_m^T A_m^{-1} w_m = \frac{\det(A_m[j])}{\det(A_m)}$ and analyzing the asymptotic behavior of numerators and denominators to establish convergence.

3. Noise Handling

Lemma 3.2 addresses the impact of additional noise terms $D^{(m)} = o(1/m)$, ensuring that main asymptotic terms remain unaffected.

Experimental Setup

Application Scenario: Continuous Lyapunov Equation

Consider Gaussian distributions satisfying the continuous Lyapunov equation: $MΣ + ΣM^T + 2I_p = 0$ where $M$ is a stable matrix whose zero pattern encodes the graph structure.

Concrete Examples

Example 2.1: Trek $T_4: 1 ← 2 → 3 → 4$

Construct the stable matrix: $M_4 = \begin{pmatrix} -1 & 1 & 0 & 0 \\ 0 & -1 & 0 & 0 \\ 0 & 1 & -1 & 0 \\ 0 & 0 & 1 & -1 \end{pmatrix}$

Obtaining the solution: $Σ_4 = \begin{pmatrix} 3/2 & 1/2 & 1/2 & 3/8 \\ 1/2 & 1 & 1/2 & 1/4 \\ 1/2 & 1/2 & 3/2 & 7/8 \\ 3/8 & 1/4 & 7/8 & 15/8 \end{pmatrix}$

Verifying $(Σ_4)_{14|23} = 1/10 ≠ 0$.

Experimental Results

Toeplitz Matrix Application

For $n$-dimensional symmetric Toeplitz matrices: $T_n = (-|i-j|)_{i,j=1,\ldots,n}$

Lemma 4.1 proves that for specific matrices $U$ and $G$: $f(U,G) = 2^{k+1-|\{1,n_l\}|-|\{1,n_r\}|} ≠ 0$

Main Results Verification

Proposition 4.2 (Proposition 4.5 of Boege et al. 2025): For any trek of arbitrary length $T: 1 ← ··· → p$, there exists a covariance matrix $Σ$ satisfying the continuous Lyapunov equation such that $X_1 \not\perp X_p | X_2, \ldots, X_{p-1}$.

The proof constructs a matrix sequence $M^{(m)}$ such that the corresponding solutions satisfy the conditions of Theorem 1.2, thereby ensuring that conditional independence does not hold.

Related Work

Theoretical Foundations

1. Stewart (1969), Ben-Israel and Greville (2003): Continuity of pseudoinverses for constant-rank matrix sequences
2. Lauritzen (1996, 2024): General theory of convergence of conditional distributions
3. Bérczi et al. (2023): This work generalizes results from Section 7

Application Background

1. Dettling et al. (2023), Varando and Hansen (2020): Continuous Lyapunov graphical models
2. Boege et al. (2025): Characterization of conditional independence in graphical models

Conclusions and Discussion

Main Conclusions

1. Under specific conditions, Gaussian conditional distributions can converge even when the limiting covariance matrix is singular
2. Condition (i) ensures that the denominator of the Schur complement does not tend to zero, while condition (ii) guarantees compatibility of image spaces
3. Theoretical results are successfully applied to conditional independence analysis in continuous Lyapunov graphical models

Limitations

1. Restrictive Conditions: Requires satisfaction of specific matrix structure conditions
2. Computational Complexity: Calculation of function $f$ involves multiple summations, potentially difficult for large matrices
3. Scope of Application: Primarily focused on Gaussian distributions; extensibility to other distribution families remains unknown

Future Directions

1. Relaxing Conditions: Seeking more general convergence conditions
2. Computational Optimization: Developing more efficient algorithms for condition verification
3. Non-Gaussian Extensions: Generalizing results to other distribution families

In-Depth Evaluation

Strengths

1. Theoretical Rigor: Provides a complete mathematical proof framework, from generalization of the matrix determinant lemma to the main theorem
2. Practical Value: Addresses real problems in graphical models, particularly applications of continuous Lyapunov equations
3. Methodological Innovation: Cleverly combines matrix analysis, probability theory, and graph theory
4. Clear Presentation: Well-structured paper with abundant examples facilitating comprehension

Weaknesses

1. Condition Complexity: Verification of theorem conditions may be difficult, particularly computation of function $f$
2. Application Limitations: Primarily applicable to specific matrix structures; generality remains to be improved
3. Numerical Stability: For nearly singular matrices, numerical computation may encounter stability issues

Impact

1. Theoretical Contribution: Provides new theoretical tools for convergence of conditional distributions under singular limits
2. Application Prospects: Potential applications in graphical models, time series analysis, and related fields
3. Methodological Inspiration: Offers new approaches for handling similar matrix limit problems

Applicable Scenarios

1. Graphical Model Analysis: Particularly models defined by continuous Lyapunov equations
2. Time Series: Multivariate time series with specific covariance structures
3. Statistical Inference: Statistical problems requiring analysis of limiting behavior of conditional distributions

References

The paper cites multiple important related works, including:

• Ben-Israel and Greville (2003): Generalized inverse theory
• Lauritzen (1996, 2024): Graphical models and conditional distribution theory
• Boege et al. (2025): Applications of conditional independence in diffusion processes
• Stewart (1969): Continuity theory of generalized inverses

This paper makes significant contributions at both theoretical and applied levels, providing new mathematical tools for understanding conditional distribution behavior under singular limits, with successful applications to conditional independence analysis in graphical models.