q-Paths: Generalizing the geometric annealing path using power means

Vaden Masrani, Rob Brekelmans, Thang Bui, Frank Nielsen, Aram Galstyan, Greg Ver Steeg, Frank Wood

Abstract

Annealing paths are essential for bridging simple base distributions and complex target distributions in statistical inference. While the geometric path is the standard choice, it often fails to efficiently navigate distribution spaces. This paper introduces q-paths, a family of paths derived from power means and nonextensive thermodynamics. By generalizing the geometric path, q-paths offer improved flexibility and performance. Empirical results in Sequential Monte Carlo and Annealed Importance Sampling demonstrate that q-paths achieve more accurate marginal likelihood estimates and tighter lower bounds for generative models.

Introduction

In many statistical applications, bridging the gap between a simple "base" distribution (e.g., a prior) and a complex "target" distribution (e.g., a posterior) is necessary for tasks like sampling or marginal likelihood estimation. This is typically achieved through annealing, creating a sequence of intermediate densities. The industry standard is the geometric path, which uses a weighted geometric average. However, the geometric path is not always the most efficient way to traverse "distribution space," particularly when dealing with multimodal targets.

The q-Path Framework

The authors propose q-paths, a novel generalization of the geometric path using the concept of generalized means (or Kolmogorov means) and "deformed" functions from nonextensive thermodynamics. The q-path formula is defined as:

$$\tilde{\pi}_{\beta,q}(z) = \left[ (1 - \beta) \pi_0(z)^{1-q} + \beta \tilde{\pi}_1(z)^{1-q} \right]^{\frac{1}{1-q}}$$

This framework provides a continuous interpolation between different types of averages. As $q \to 1$, the formula recovers the standard geometric path. When $q = 0$, it becomes the arithmetic mean (a mixture distribution).

Visualizing the transition between arithmetic ($q=0$) and geometric ($q=1$) paths.

Unlike previous alternatives like the moment-averaging path, q-paths have a simple analytical form. This allows them to be directly used as energy functions in MCMC samplers such as Hamiltonian Monte Carlo (HMC) without the need for complex approximations.

Key Insights: Mass-Covering and Stability

A significant advantage of q-paths is their "mass-covering" property. While the geometric path ($q=1$) can be "aggressive" and potentially miss important regions of probability mass, setting $q$ slightly less than 1 (e.g., $q=0.99$) makes the intermediate densities behave more like mixture models. This helps "cover" the space between distributions more effectively during the annealing process.

The paper also connects Bayesian inference with Tsallis statistics, showing that "deformed" functions used in physics for complex systems are equally applicable to navigating complex probability landscapes. While $q$ values far from 1 provide better mass coverage, they can introduce numerical instability. Empirically, small deviations (e.g., $q = 1 - \delta$) offer the best balance between performance gains and stability.

Empirical Results

The authors evaluated q-paths on several benchmarks, including Bayesian logistic regression (Pima Indians Diabetes and Sonar datasets) and Variational Autoencoder (VAE) evaluation (Omniglot dataset).

In Sequential Monte Carlo (SMC) experiments, q-paths consistently achieved lower errors and reduced variance in marginal likelihood estimation compared to the standard geometric path. For generative model evaluation using Annealed Importance Sampling (AIS), q-paths provided tighter lower bounds on log-likelihood.

The Bidirectional Monte Carlo (BDMC) gap as a function of $q$, showing an optimal performance region for $q < 1$.

As shown in the evaluation of VAEs, the Bidirectional Monte Carlo (BDMC) gap—a measure of evaluation accuracy—is minimized at values of $q$ slightly below 1, confirming that the generalized path outperforms the traditional geometric approach.

Conclusion

q-paths offer a mathematically elegant and computationally tractable alternative to the standard geometric annealing path. By leveraging power means, they provide a flexible way to handle complex distributions in Bayesian inference and generative modeling. Future work may explore their application in Parallel Tempering or the Thermodynamic Variational Objective (TVO).

Summary updated

Abstract

Many common machine learning methods involve the geometric annealing path, a sequence of intermediate densities between two distributions of interest constructed using the geometric average. While ...
BibTeX
@inproceedings{masrani2021q,
  author = {Vaden Masrani and Rob Brekelmans and Thang Bui and Frank Nielsen and Aram Galstyan and Greg Ver Steeg and Frank Wood},
  title = {{q-Paths: Generalizing the geometric annealing path using power means}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=7hRdrxbg7},
}
Figure