Asymptotics for parametric martingale posteriors

1. Problem Statement

The paper addresses the computational and theoretical limitations of traditional Bayesian inference, particularly regarding the reliance on Markov Chain Monte Carlo (MCMC) for posterior sampling and the necessity of specifying a prior distribution.

• Context: The Martingale Posterior framework (Fong et al., 2023) generalizes Bayesian inference by replacing the likelihood and prior with a sequence of one-step-ahead predictive densities. Posterior sampling is achieved via predictive resampling (sequential imputation of unseen data), which is parallelizable and avoids MCMC convergence issues.
• Gap: While recent work has applied this framework using plug-in parametric predictive densities (e.g., using Maximum Likelihood Estimates updated via stochastic gradient descent), the asymptotic properties of these parametric martingale posteriors were not rigorously established. Specifically, there was a lack of understanding regarding:
  1. The convergence behavior of the resampling scheme as the imputed population size grows (Predictive Asymptotics).
  2. The frequentist properties (consistency and coverage) as the observed sample size grows (Bernstein-von Mises properties).
  3. Efficient methods to approximate the posterior without running the resampling to infinity.

2. Methodology

The authors investigate a specific class of parametric martingale posteriors where the predictive density $p_{\theta}(y)$ belongs to a parametric family.

Core Mechanism: Recursive Update

Given observed data $Y_{1:n}$ and an initial estimator $\theta_n$ (e.g., MLE), the framework generates a sequence of imputed data points $Y_{n+1}, Y_{n+2}, \dots$ and updates the parameter estimate recursively:
$$\theta_N = \theta_{N-1} + N^{-1} I(\theta_{N-1})^{-1} s(\theta_{N-1}, Y_N)$$
where:

• $s(\theta, y) = \nabla_\theta \log p_\theta(y)$ is the score function.
• $I(\theta)$ is the Fisher information matrix.
• The term $I(\theta)^{-1}s(\theta, y)$ is the natural gradient.
• The learning rate $N^{-1}$ satisfies standard stochastic approximation conditions ($\sum N^{-1} = \infty, \sum N^{-2} < \infty$).

The parametric martingale posterior is defined as the distribution of the limit $\theta_\infty = \lim_{N \to \infty} \theta_N$ conditional on the observed data $Y_{1:n}$.

Theoretical Framework

The authors utilize Martingale Limit Theory (specifically the Martingale Central Limit Theorem) to derive two distinct asymptotic results under different regimes:

1. Predictive Asymptotics: Treating observed data $Y_{1:n}$ as fixed and letting the imputed population size $N \to \infty$.
2. Frequentist Asymptotics: Treating the observed data size $n \to \infty$ (Bernstein-von Mises regime).

3. Key Contributions

A. Predictive Central Limit Theorem (Theorem 1)

The paper establishes the limiting distribution of the difference between the limit $\theta_\infty$ and the current estimate $\theta_N$ as $N \to \infty$.

• Result: The scaled difference converges to a standard normal distribution, but with a random variance:
  $$V_N^{-1} (\theta_\infty - \theta_N) \xrightarrow{d} \mathcal{N}(0, 1)$$
  where $V_N^2 \approx I(\theta_\infty)^{-1} \sum_{i=N}^\infty i^{-2}$.
• Significance: Unlike classical CLTs where variance converges to a constant, here the variance depends on the random limit $\theta_\infty$. This result relies on stable convergence (Rényi, 1963).
• Practical Application: This theorem enables a Hybrid Predictive Resampling Algorithm. Instead of running the recursive update to a massive $N$ (which is computationally expensive), one can run it to a moderate $N$ and then approximate the remaining gap using a Gaussian draw with the estimated random variance. This drastically accelerates sampling.

B. Bernstein-von Mises (BvM) Theorem for Martingale Posteriors (Theorem 3)

The paper proves that under frequentist asymptotics ($n \to \infty$), the martingale posterior behaves like a standard Bayesian posterior.

• Result: If the initial estimator $\theta_n$ is strongly consistent and the model is well-specified, the posterior distribution centered at $\theta_n$ converges to a normal distribution:
  $$r_n^{-1}(\theta_{n\infty} - \theta_n) \xrightarrow{d} \mathcal{N}(0, I(\theta^*)^{-1})$$
  where $r_n \approx n^{-1/2}$.
• Novelty: This is the first BvM result for martingale posteriors. It confirms that the method attains desirable frequentist coverage properties (credible intervals match confidence intervals) without requiring a prior.
• Centring vs. Uncertainty: The authors highlight a unique feature of this framework: the centring (location) depends entirely on the initial estimator $\theta_n$, while the uncertainty (variance) is determined by the recursive update rule and the Fisher information. This separation allows for flexible construction of the posterior.

C. Conditions for Validity

The paper introduces Assumption 2 (Uniform Integrability of the natural gradient) and provides Proposition 1, a practical checkable condition involving the fourth moment of the natural gradient ($E[Z^4] \leq B + C\theta^4$). This ensures the martingale is bounded in $L^2$ and the CLT holds.

4. Results

Simulation Studies

• Setup: Bivariate normal model with unknown mean and covariance.
• Findings:
  • Coverage: The parametric martingale posterior achieves coverage rates comparable to traditional Bayesian inference (using MCMC) as $n$ increases. For small $n$, it slightly undercovers if the initial estimator is poor, but matches the Bayesian posterior for $n \geq 100$.
  • Speed: The martingale posterior is orders of magnitude faster than MCMC. For $n=500$, MCMC took ~14.7s per sample, while the martingale posterior took ~0.003s.
  • Hybrid Efficiency: The hybrid sampling scheme (truncating at $N=n+50$ and adding Gaussian noise) produced results indistinguishable from the "exact" scheme (truncating at $N=n+50000$) but with negligible computational cost.

Real Data Example (AIDS Clinical Trials)

• Task: Robust regression on CD4 counts using a heavy-tailed Student-t model.
• Findings:
  • The hybrid martingale posterior produced 95% credible intervals and coefficient estimates nearly identical to the Bayesian posterior (with weakly informative priors).
  • Computation: The Bayesian MCMC approach took 132.4s, while the martingale posterior (hybrid) took 0.03s. The exact truncation ($N=n+50000$) took 14.4s.
  • The study demonstrated that the martingale posterior is a viable "frequentist" alternative that requires no prior elicitation.

5. Significance and Implications

1. Computational Efficiency: The paper provides a theoretical justification for a hybrid sampling algorithm that bypasses the linear computational cost of predictive resampling. By leveraging the predictive CLT, one can approximate the infinite imputation limit with a single Gaussian step, making the method scalable to large datasets.
2. Prior-Free Inference: The framework offers a rigorous path to Bayesian-like inference without specifying a prior distribution, relying instead on predictive modeling and stochastic approximation.
3. Frequentist Guarantees: The BvM theorem ensures that the method is not just a heuristic but possesses strong frequentist properties (consistency and correct asymptotic variance), bridging the gap between Bayesian and frequentist paradigms.
4. Flexibility: The separation of the initial estimator (centring) and the update rule (uncertainty) allows practitioners to use efficient point estimators (like MLE) while still obtaining a full posterior distribution with correct uncertainty quantification.
5. Extensions: The methodology is extended to multivariate parameters and regression settings (including fixed and random designs), demonstrating broad applicability beyond simple i.i.d. models.

In summary, this paper solidifies the theoretical foundation of parametric martingale posteriors, proving they are asymptotically normal and frequentist-valid, while providing a practical algorithm that achieves massive computational speedups over traditional MCMC methods.