A note on target distribution ambiguity of likelihood-free samplers

Scott A. Sisson, Gareth W. Peters, Mark Briers, Yanan FanarXiv 2010 · DOI · arXiv

Abstract

Methods for Bayesian simulation in the presence of computationally intractable likelihood functions are of growing interest. Termed likelihood-free samplers, standard simulation algorithms such as Markov chain Monte Carlo have been adapted for this setting. In this article, by presenting generalisations of existing algorithms, we demonstrate that likelihood-free samplers can be ambiguous over the form of the target distribution. We also consider the theoretical justification of these samplers. Distinguishing between the forms of the target distribution may have implications for the future development of likelihood-free samplers.

Introduction

In Bayesian inference, the posterior distribution is typically derived using a likelihood function—the probability of observing specific data given a set of parameters. However, in complex scientific domains such as population genetics, epidemiology, and climate modeling, the likelihood function is often "computationally intractable." While researchers can easily simulate data from a model, they cannot explicitly calculate the mathematical probability of that data.

To circumvent this, likelihood-free samplers, commonly known as Approximate Bayesian Computation (ABC), have become essential. These methods replace the evaluation of the likelihood with a simulation step: if a simulated dataset is sufficiently similar to the observed data, the parameters that generated it are accepted. While these methods are widely used, a fundamental theoretical question has persisted: what exactly are these algorithms sampling? This paper explores the "target distribution ambiguity" between marginal and joint posterior interpretations.

The Likelihood-Free Framework

The authors examine three generalized likelihood-free algorithms, extended to allow for $S \ge 1$ simulated datasets per iteration:

  1. LF-REJ (Rejection Sampling): The most basic form, where parameters are drawn from the prior and accepted or rejected based on the proximity of simulated data to observed data.
  2. LF-MCMC (Markov chain Monte Carlo): An iterative approach that explores the parameter space via a random walk, using simulation-based acceptance ratios.
  3. LF-SMC (Sequential Monte Carlo): A population-based method that evolves a set of "particles" through successive distributions to identify high-probability regions.

The core of the investigation lies in how we interpret the acceptance probabilities (in MCMC) and importance weights (in SMC) within these algorithms.

Target Distribution Ambiguity

The paper identifies two distinct ways to interpret the target of these samplers:

  • The Marginal Target: The goal is to estimate the posterior of the parameters $\theta$ directly. In this view, the simulation step is merely a Monte Carlo approximation of an integral.
  • The Joint Target: The goal is to estimate the posterior of both the parameters $\theta$ and the auxiliary simulated data $t$ simultaneously. Here, the simulated data is treated as a set of auxiliary variables within an augmented model.

The "Ratio Bias" Trap

If one adopts the marginal perspective, a significant theoretical problem arises. In LF-MCMC, the acceptance ratio involves a ratio of two Monte Carlo estimates. Mathematically, the expectation of a ratio is not equal to the ratio of the expectations ($E[X/Y] \neq E[X]/E[Y]$). This implies that for any finite number of simulations $S$, the algorithm should technically be biased, only becoming exact as $S$ approaches infinity.

The Joint Distribution Rescue

The authors demonstrate that this perceived bias is an artifact of the marginal interpretation. If the algorithm is instead viewed as targeting the joint distribution of parameters and auxiliary data, the mathematical "bias" disappears. Under the joint interpretation, the algorithms are exact for any $S \ge 1$. This provides a vital theoretical safety net: the actual code executed by a researcher is identical regardless of the interpretation, but the joint view provides the rigorous justification that the marginal view lacks.

Key Findings

  • Rejection Sampling Robustness: The LF-REJ algorithm is unique in that it remains unbiased under both the marginal and joint interpretations for any number of simulations.
  • Validation of MCMC and SMC: While LF-MCMC and LF-SMC appear biased from a marginal standpoint, they are perfectly valid and exact when viewed as joint space samplers. This confirms that practitioners can rely on these methods even with a small number of simulations per step.
  • Theoretical Equivalence: The "ambiguity" is conceptual rather than functional. The simulated datasets act as auxiliary parameters in the joint view, whereas they are treated as noise in the marginal view.

Conclusion and Future Implications

This research offers a "get out of jail free" card for current practitioners of ABC, proving that their methods are theoretically sound when viewed through the lens of joint distribution sampling. However, the authors issue a warning for the future. The current theoretical harmony relies on the fact that intractable likelihood terms cancel out in the acceptance ratios of MCMC and SMC.

If future, more complex likelihood-free algorithms are developed where these terms do not cancel, the joint distribution interpretation may no longer be applicable. In such cases, researchers would be forced back into the marginal approach, where the ratio bias becomes a real and non-negligible obstacle. The paper concludes by urging the community to be explicit about their target distributions to avoid future theoretical pitfalls as the field evolves.

Summary updated

Abstract

Methods for Bayesian simulation in the presence of computationally intractable likelihood functions are of growing interest. Termed likelihood-free samplers, standard simulation algorithms such as Markov chain Monte Carlo have been adapted for this setting. In this article, by presenting generalisations of existing algorithms, we demonstrate that likelihood-free samplers can be ambiguous over the form of the target distribution. We also consider the theoretical justification of these samplers. Distinguishing between the forms of the target distribution may have implications for the future development of likelihood-free samplers.
BibTeX
@misc{sisson2010note,
  author = {Scott A. Sisson and Gareth W. Peters and Mark Briers and Yanan Fan},
  title = {{A note on target distribution ambiguity of likelihood-free samplers}},
  year = {2010},
  eprint = {1005.5201},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1005.5201},
}
Figure