Introduction
In Markov chain Monte Carlo (MCMC) theory, the gold standard for convergence is geometric ergodicity, where a chain converges to its target distribution at an exponential rate. Traditionally, researchers establish this property using functional analysis—specifically by identifying a spectral gap—or through the construction of Lyapunov drift conditions. However, many modern Bayesian computational methods, such as pseudo-marginal MCMC and Approximate Bayesian Computation (ABC), do not satisfy these stringent requirements. Instead, they exhibit subgeometric convergence, where the rate of decay is polynomial or log-polynomial.
Analyzing subgeometric chains has historically been a cumbersome process, often requiring the construction of complex, case-specific Lyapunov functions. This paper introduces a more general and "user-friendly" functional-analytic toolkit based on Weak Poincaré Inequalities (WPI) to characterize these slower convergence rates without the need for traditional drift conditions.
The Weak Poincaré Inequality Framework
The core innovation of this work lies in the transition from strong to weak Poincaré inequalities. While a strong Poincaré inequality implies a constant spectral gap and thus exponential convergence, a WPI allows the "gap" to shrink as one explores different regions of the function space. This flexibility is essential for capturing the behavior of chains that lack a uniform spectral gap.
The authors provide two primary parameterizations for WPIs—denoted as $\alpha$ and $\beta$—which allow the framework to adapt to various noise distributions and target geometries. A central methodological contribution is Theorem 8, which provides a direct quantitative mapping from a WPI to a convergence bound for the Markov chain. This allows researchers to translate functional-analytic properties into practical convergence guarantees.
Comparison and Chaining Techniques
A significant challenge in pseudo-marginal MCMC is that researchers often only have access to noisy estimates of the target density. The authors address this by developing "chaining" or comparison theorems. These theorems allow a researcher to inherit convergence properties from a well-understood "marginal" chain and apply them to a more complex "pseudo-marginal" version, provided the two are sufficiently "close."
This approach clarifies the cost of noise. The paper precisely characterizes how the convergence rate degrades when an exact likelihood is replaced by a noisy estimate, linking this degradation directly to the tail behavior of the noise weights. Interestingly, while averaging multiple noisy estimates (to reduce variance) improves the constant factors and short-term speed, the authors demonstrate that it does not necessarily alter the fundamental subgeometric rate of convergence.
Applications and Key Results
The utility of the WPI framework is demonstrated across several algorithmic scenarios:
- Independent Metropolis-Hastings (IMH): The framework successfully recovered known polynomial convergence rates for IMH samplers, validating the WPI approach against established results that were previously much harder to derive.
- Approximate Bayesian Computation (ABC): ABC algorithms often suffer from subgeometric convergence due to the use of local proposals. The authors provide a rigorous theoretical safety net by proving that ABC convergence can be bounded using WPI, offering new reliability for ABC users.
- Optimal Tuning for PMMH: For the widely used Particle Marginal Metropolis-Hastings (PMMH) algorithm with lognormal noise, the authors derived a specific tuning target. They found that the optimal noise variance $\sigma$ is approximately 0.973. This provides a concrete guideline for practitioners to minimize the bound on asymptotic variance.
Discussion and Limitations
While the WPI framework offers a powerful alternative to Lyapunov-based analysis, it is not without limitations. The current theory relies heavily on the reversibility of the Markov chain (detailed balance). Extending these results to non-reversible algorithms, such as Piecewise Deterministic Markov Processes (PDMPs), remains an open area of research.
Additionally, while WPI-based bounds are mathematically rigorous, they may lose some precision compared to direct spectral analysis in the limit where a strong Poincaré inequality holds. Finally, the scaling of these bounds in high-dimensional state spaces is a critical area for future investigation to ensure the framework remains robust as problem complexity increases.
Conclusion
By leveraging Weak Poincaré Inequalities, this work provides a transparent and elegant method for analyzing the convergence of subgeometric Markov chains. It bridges the gap between complex drift conditions and practical MCMC diagnostics, offering both theoretical insights into the nature of noise in Bayesian computation and practical guidelines for algorithm tuning.
Summary updated