Structured variational inference in Bayesian state-space models

Honggang Wang, Anirban Bhattacharya, Debdeep Pati, Yun YangAISTATS 2022 · OpenReview

Abstract

Variational inference is routinely deployed in Bayesian state-space models as an efficient computational technique. Motivated by the inconsistency issue observed by Wang and Titterington (2004) for the mean-field approximation in linear state-space models, we consider a more expressive variational family for approximating the joint posterior of the latent variables to retain their dependence, while maintaining the mean-field (i.e. independence) structure between latent variables and parameters. In state-space models, such a latent structure adapted mean-field approximation can be efficiently computed using the belief propagation algorithm. Theoretically, we show that this adapted mean-field approximation achieves consistency of the variational estimates. Furthermore, we derive a non-asymptotic risk bound for an averaged alpha-divergence from the true data generating model, suggesting that the posterior mean of the best variational approximation for the static parameters shows optimal concentration. From a broader perspective, we add to the growing literature on statistical accuracy of variational approximations by allowing dependence between the latent variables, and the techniques developed here should be useful in related contexts.

Introduction

Bayesian State-Space Models (SSMs) are fundamental to time-series analysis, providing a robust framework for tracking objects, modeling financial markets, and understanding biological signals. However, the exact posterior distribution for these models is often computationally intractable due to the complex dependencies between latent states and observed data.

To address this, researchers frequently employ Variational Inference (VI), which approximates the posterior with a simpler, more tractable distribution. The most common approach is the mean-field approximation, which assumes all variables are mutually independent. While computationally efficient, this assumption is fundamentally flawed for SSMs. As demonstrated by Wang and Titterington (2004), mean-field VI produces inconsistent estimates for emission parameters in linear Gaussian models because it ignores the inherent temporal dependencies of the hidden states. This paper provides a rigorous theoretical foundation for "structured" VI, demonstrating that preserving latent dependencies restores consistency and achieves near-optimal convergence rates.

The Structured Variational Approach

The authors advocate for a Generalized Mean-Field (GMF) inference strategy. Rather than assuming total independence, they propose a strategic "block" decomposition:

  1. Latent Block: The hidden states ($X_1, \dots, X_n$) are treated as a single joint block, preserving their Markovian dependence.
  2. Parameter Block: Static parameters ($\theta$) are kept independent of the latent states.

To ensure mathematical robustness, the study utilizes the $\alpha$-Variational Bayes ($\alpha$-VB) framework. By "tempering" the likelihood (raising it to a power $\alpha \in (0, 1)$), the authors simplify the theoretical analysis of risk bounds—measuring the distance between the estimate and the ground truth—without compromising practical utility.

Computational Implementation

The implementation relies on the Coordinate Ascent Variational Inference (CAVI) algorithm. The critical challenge of updating the latent block is addressed via the Belief Propagation (BP) algorithm, specifically the Forward-Backward or Rauch–Tung–Striebel smoother. This allows the algorithm to handle the chain-like dependencies of the hidden states in $O(n)$ time, maintaining the efficiency that makes VI attractive in the first place.

Theoretical Insights and Key Results

1. Consistency Restored

The paper’s primary contribution is the formal proof (Theorem 2) that the failure of standard VI is a direct consequence of the mean-field factorization. By coupling the latent states, the structured approach eliminates the systematic bias found in standard mean-field methods, leading to consistent parameter estimates.

2. Near-Optimal Convergence Rates

For stable Linear Gaussian SSMs, the authors demonstrate that the variational posterior concentrates around the true parameters at a rate of approximately $O(1/n)$ (ignoring logarithmic factors). This matches the "minimax" optimal rate, indicating that the structured VI approach is as theoretically efficient as exact inference in this context. In multivariate cases, the error scales with $d^2/n$ (where $d$ is the state dimension), aligning with standard parametric estimation expectations.

3. The Role of Stability and "Blow-up" Functions

The quality of inference is intrinsically linked to the stability of the system. In unstable models (e.g., where transition matrix eigenvalues exceed 1), latent variables do not "mix" well, causing error bounds to grow exponentially. To account for how dependencies propagate over time, the authors introduced "blow-up functions" ($f_\lambda(n)$ and $f_\mu(n)$), providing the first non-asymptotic risk bounds for these complex models.

Empirical Validation

While primarily theoretical, the findings were validated using simulated Scalar and Multivariate Linear Gaussian State-Space Models. These simulations confirmed that as iterations increase, the Bayes risk and Root Mean Square Error (RMSE) for the structured approach converge toward zero. In contrast, the standard mean-field approach remained biased regardless of the data volume.

Conclusion and Future Directions

This work bridges the gap between the practical necessity of VI and the theoretical requirement for statistical accuracy in time-series modeling. By allowing dependence within the latent block, the authors provide a template for more accurate inference in structured models.

Future research may expand these proofs to:

  • Time-Inhomogeneous Models: Addressing systems where parameters fluctuate over time.
  • Complex Topologies: Extending the theory from simple chains to trees or loopy graphs.
  • Algorithmic Convergence: Investigating the specific number of CAVI iterations required to reach optimal solutions under varying conditions.

Summary updated

Abstract

Variational inference is routinely deployed in Bayesian state-space models as an efficient computational technique. Motivated by the inconsistency issue observed by Wang and Titterington (2004) for the mean-field approximation in linear state-space models, we consider a more expressive variational family for approximating the joint posterior of the latent variables to retain their dependence, while maintaining the mean-field (i.e. independence) structure between latent variables and parameters. In state-space models, such a latent structure adapted mean-field approximation can be efficiently computed using the belief propagation algorithm. Theoretically, we show that this adapted mean-field approximation achieves consistency of the variational estimates. Furthermore, we derive a non-asymptotic risk bound for an averaged alpha-divergence from the true data generating model, suggesting that the posterior mean of the best variational approximation for the static parameters shows optimal concentration. From a broader perspective, we add to the growing literature on statistical accuracy of variational approximations by allowing dependence between the latent variables, and the techniques developed here should be useful in related contexts.
BibTeX
@inproceedings{wang2022structured,
  author = {Honggang Wang and Anirban Bhattacharya and Debdeep Pati and Yun Yang},
  title = {{Structured variational inference in Bayesian state-space models}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
  url = {https://openreview.net/forum?id=6X5yLq1iA9w},
}

Loading related content...

Figure