Martingale posterior distributions

Edwin Fong, Chris Holmes, Stephen G. WalkerarXiv 2021 · DOI · arXiv

TL;DR

This paper introduces the Martingale Posterior, a framework that performs Bayesian uncertainty quantification by focusing on "missing observations" rather than abstract parameters. By recursively sampling from a predictive distribution—a process called predictive resampling—researchers can generate posterior samples for any statistic without ever explicitly defining a likelihood or a prior, offering a computationally efficient and GPU-friendly alternative to MCMC.

What Problem Does This Paper Address?

In traditional Bayesian statistics, you start with a prior (what you believe before seeing data) and a likelihood (how the data is generated). You then use Bayes' Rule to find the posterior, which tells you how much to update your beliefs.

However, this approach has two major friction points:

  1. Metaphysical Parameters: Eliciting a prior on an abstract mathematical parameter (like a "precision matrix") is often unintuitive for practitioners.
  2. Computational Bottlenecks: For complex models, calculating the posterior requires Markov Chain Monte Carlo (MCMC) sampling, which is notoriously slow, difficult to parallelize, and struggles with high-dimensional data.

The authors argue that statistical uncertainty actually stems from missing data. If we observed the entire infinite population, we would know any parameter of interest perfectly. Therefore, the "true" job of a Bayesian is to predict those missing observations.

Methods

The authors propose a "predictive-first" approach. Instead of a prior-to-posterior pipeline, they use Predictive Resampling:

  1. Start with observed data $y_{1:n}$.
  2. Impute the next observation $Y_{n+1}$ by sampling from a 1-step-ahead predictive distribution $p(y | y_{1:n})$.
  3. Update the predictive model using this new "synthetic" data point.
  4. Repeat until you have a large "imputed population."
  5. Calculate the statistic (e.g., the mean, median, or number of modes) on this full population.

To make this practical for multivariate data, regression, and classification, the authors utilize Bivariate Copulas. They build recursive updates where the predictive density at step $i+1$ is a mixture of the previous density and a "copula kernel" centered at the new observation.

Crucially, they rely on the theory of Martingales and Conditionally Identically Distributed (c.i.d.) sequences. This ensures that the sequence of predictive distributions converges to a valid random limit, which they term the Martingale Posterior.

Datasets

The authors evaluate their method against standard benchmarks like Dirichlet Process Mixture Models (DPMM) and Gaussian Processes (GP) using:

  • Simulated Data: Univariate and high-dimensional Gaussian Mixture Models (GMM), and the non-linear "Moon" classification dataset.
  • Density Estimation: The classic Galaxy dataset (velocity of galaxies) and the Air Quality dataset (ozone and solar radiation).
  • Multivariate UCI Benchmarks: Breast Cancer (26D), Ionosphere (32D), Parkinsons (16D), and Wine (13D).
  • Regression: The LIDAR dataset (light detection and ranging) for non-linear, heteroscedastic regression.
  • High-Dimensional Regression/Classification: UCI datasets including Boston Housing, Concrete, and Statlog.

Key Insights and Ideas

  • The "Aha" Moment: Bayesian uncertainty is just the distribution of a statistic over all possible ways the "missing" part of the population could turn out.
  • Generalizing the Bootstrap: The well-known Bayesian Bootstrap is actually a special, "discrete" case of the Martingale Posterior where the predictive distribution is simply the empirical distribution of the data.
  • No MCMC Required: Because the copula updates are recursive and exact, the method is "GPU-friendly" and can be parallelized across different posterior samples.
  • Prior Elicitation via Observables: Instead of a prior on $\theta$, you provide an "initial guess" for the predictive density $p_0(y)$, which is much easier to visualize and justify.

Key Results

  • Speed: In the Galaxy dataset, predictive resampling took 2 seconds for 1,000 samples, compared to 25 seconds for MCMC-based DPMM. For the Air Quality dataset, the gap was even larger: 10 seconds vs. 4 minutes.
  • Accuracy: In multivariate density estimation (Table 1), the Copula-based Martingale Posterior consistently outperformed Kernel Density Estimation (KDE) and Variational Inference (VI) for DPMMs in terms of test log-likelihood.
  • Scalability: Unlike Variational Inference, which degrades quickly as dimensionality $d$ increases, the copula method remained robust in high-dimensional settings (up to $d=100$ in simulations).
  • Theoretical Soundness: The authors proved frequentist consistency for their multivariate copula density estimator, showing it converges to the true distribution as the sample size $n$ grows.

Limitations and Future Work

  • Permutation Dependence: The recursive update depends on the order in which data is processed. While the authors mitigate this by averaging over 10 random permutations, it remains a theoretical departure from strict exchangeability.
  • Out-of-Distribution Behavior: In regression tasks, the authors noted that uncertainty can sometimes "collapse" to zero for $x$ values far away from the observed data if the resampling scheme for covariates is too simple.
  • Complex Data Structures: The current work focuses on i.i.d. (independent and identically distributed) data. Extending this to time series, spatial data, or hierarchical models is the next major frontier for Martingale Posteriors.

Summary updated

Abstract

The prior distribution on parameters of a sampling distribution is the usual starting point for Bayesian uncertainty quantification. In this paper, we present a different perspective which focuses on missing observations as the source of statistical uncertainty, with the parameter of interest being known precisely given the entire population. We argue that the foundation of Bayesian inference is to assign a distribution on missing observations conditional on what has been observed. In the conditionally i.i.d. setting with an observed sample of size $n$, the Bayesian would thus assign a predictive distribution on the missing $Y_{n+1:\infty}$ conditional on $Y_{1:n}$, which then induces a distribution on the parameter. Demonstrating an application of martingales, Doob shows that choosing the Bayesian predictive distribution returns the conventional posterior as the distribution of the parameter. Taking this as our cue, we relax the predictive machine, avoiding the need for the predictive to be derived solely from the usual prior to posterior to predictive density formula. We introduce the \textit{martingale posterior distribution}, which returns Bayesian uncertainty directly on any statistic of interest without the need for the likelihood and prior, and this distribution can be sampled through a computational scheme we name \textit{predictive resampling}. To that end, we introduce new predictive methodologies for multivariate density estimation, regression and classification that build upon recent work on bivariate copulas.
BibTeX
@misc{fong2021martingale,
  author = {Edwin Fong and Chris Holmes and Stephen G. Walker},
  title = {{Martingale posterior distributions}},
  year = {2021},
  eprint = {2103.15671},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.15671},
}

Loading related content...

Figure