Martingale Posterior Framework

• Martingale Posterior Framework is a generalization of Bayesian inference that uses sequential one-step-ahead predictive densities to construct posterior distributions via forward imputation.
• It leverages martingale and exchangeability properties to ensure coherent one-step predictive performance while highlighting challenges in determining multi-step predictions.
• Its computational strategies, including predictive resampling and score-driven updates, avoid MCMC and enhance efficiency in uncertainty quantification across diverse applications.

Searching arXiv for the specified paper and closely related martingale-posterior work to ground the article in current literature. arXiv search query: all:"martingale posterior" OR ti:"Martingale Posterior" OR ([2510.25154](/papers/2510.25154)) The martingale posterior framework is a generalization of Bayesian inference in which one elicits a sequence of one-step-ahead predictive densities instead of the likelihood and prior, and posterior sampling then involves the imputation of unseen observables through predictive resampling rather than Markov chain Monte Carlo. In this formulation, statistical uncertainty is attributed to missing observations, with the parameter or functional of interest becoming known precisely given the entire population. The framework was introduced as a predictive alternative to prior–likelihood specification and has since been developed for density estimation, regression, classification, quantile inference, foundation-model-based tabular uncertainty quantification, neural processes, deep neural networks, discretely observed diffusions, and federated learning (Fong et al., 2021).

1. Formal definition and predictive-Bayes formulation

A standard formulation considers an infinite sequence of observations $(Z_i)_{i\ge1}$ in some space $\mathcal Z$. A predictive rule is a sequence of one-step-ahead conditional distributions

$$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$

Writing $p_i(z)$ for the density or mass function of $P_i$, and defining the empirical distribution

$F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$

the martingale posterior for a functional $\theta=\Theta(F)$ is the conditional law of $\theta(F_\infty)$ given the observed data $z_{1:n}$, where $F_N\to F_\infty$ almost surely under forward sampling induced by the predictive rule (Ng et al., 29 Oct 2025).

This predictive specification replaces the prior–likelihood pair in classical Bayes. If the predictive rule coincides with the true Bayesian posterior predictive distribution, the martingale posterior recovers the classical posterior exactly. In the original formulation, the framework is expressed as a posterior on a statistic of the infinite empirical distribution, with posterior draws obtained by augmenting the observed sample with forward-simulated future observations and evaluating the statistic on the augmented empirical distribution (Fong et al., 2021).

The induced posterior on $\mathcal Z$0 is written as

$\mathcal Z$1

and in practice it is approximated by repeated predictive-resampling paths. Starting from $\mathcal Z$2, one sequentially draws $\mathcal Z$3 according to the specified one-step predictives, computes $\mathcal Z$4, and uses the empirical distribution of $\mathcal Z$5 as an approximation to the martingale posterior (Ng et al., 29 Oct 2025).

2. Martingale conditions, exchangeability, and what they do—and do not—guarantee

A convenient sufficient condition for the almost sure convergence $\mathcal Z$6 is that the predictive sequence itself form a martingale: for every measurable set $\mathcal Z$7,

$\mathcal Z$8

The literature also records a weaker almost c.i.d. condition,

$\mathcal Z$9

as a sufficient guarantee for existence of the limit random measure. In supervised settings, permutation invariance is often desired so that the predictive rule depends only on the multiset of past observations (Ng et al., 29 Oct 2025).

The foundational theory emphasizes conditionally identically distributed predictives and exchangeability. For each fixed $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$0, the predictive CDF sequence satisfies

$$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$1

which yields a random limit $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$2 by Doob’s martingale convergence theorem. The induced posterior on the target functional is then coherent in the sense that present predictive summaries are conditional expectations of their limiting counterparts (Fong et al., 2021).

An important clarification is that one-step coherence does not, in general, determine multi-step predictive structure. For exchangeable Bernoulli sequences, the $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$3-step all-zeros predictive is

$$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$4

which depends on all posterior moments up to order $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$5 by binomial expansion. The first moment alone is not sufficient to uniquely identify these quantities: for $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$6, the mapping from posterior mean to $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$7-step predictive is set-valued. The closure theorem states that a martingale posterior determines all $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$8-step predictives if and only if the conditional law of the terminal value is uniquely specified, equivalently if all posterior moments are unique. Hill’s $$P_0(\cdot)=\Pr(Z_1\in\cdot),\qquad P_i(\cdot\mid z_{1:i})=\Pr(Z_{i+1}\in\cdot\mid Z_{1:i}=z_{1:i}),\; i\ge1.$$9 rule under the Jeffreys $p_i(z)$0 prior is given as a positive example, and the discrepancy between plug-in and Bayes predictives is $p_i(z)$1, vanishing as the posterior concentrates (Polson et al., 28 Feb 2026).

This resolves a common misconception: the martingale posterior framework guarantees one-step predictive coherence under its core conditions, but predictive completeness for longer blocks requires stronger structural information than the first conditional moment alone (Polson et al., 28 Feb 2026).

3. Predictive resampling and score-driven computation

The generic computational mechanism is predictive resampling. A finite approximation draws $p_i(z)$2 independent predictive-resampling paths, each obtained by sequentially simulating future observations from the predictive rule and recomputing the target functional on the empirical distribution of observed plus imputed data. This produces independent posterior draws and avoids MCMC entirely (Ng et al., 29 Oct 2025).

A parametric and especially tractable variant uses score functions. For a parametric family $p_i(z)$3 with score

$p_i(z)$4

one initializes $p_i(z)$5 and updates for $p_i(z)$6 by

$p_i(z)$7

Because $p_i(z)$8, the parameter sequence is a martingale, and under regularity conditions it converges almost surely to $p_i(z)$9. The method requires gradients from a parametric density family, does not rely on MCMC algorithms, and can be implemented in parallel (Cui et al., 3 Jan 2025).

The deep-learning exposition adopts the same principle with the stochastic-gradient-type recursion

$P_i$0

where $P_i$1 and $P_i$2. Under boundedness and integrability assumptions, $P_i$3 is a martingale and converges almost surely to a limiting random variable $P_i$4; repeated independent runs then approximate the martingale posterior over network parameters (Zhumekenov et al., 14 Jun 2026).

Asymptotic theory for parametric martingale posteriors adds two central approximations. A predictive central limit theorem gives a Gaussian approximation for the un-simulated tail $P_i$5, supporting a hybrid sampling algorithm in which one simulates only up to a moderate truncation $P_i$6 and then adds a normal correction. A Bernstein–von Mises result shows that, with natural-gradient preconditioning, the martingale posterior is asymptotically

$P_i$7

so that asymptotically its credible intervals coincide with frequentist confidence intervals (Fong et al., 2024).

4. Major constructions and domain-specific variants

The framework has been instantiated through several distinct predictive mechanisms. The following constructions are explicitly developed in the literature.

Construction	Core mechanism	Domain
Copula-based martingale posterior	Bivariate copula recursions satisfying the martingale condition	Density estimation, regression, classification
Moment martingale posterior	Mixture predictive combining parametric and nonparametric components with moment-martingale updates	Semiparametric predictive Bayes
Quantile martingale posterior	Recursive updates of a smooth quantile function with increasing rearrangement	Quantile estimation and regression
PFN-based martingale posterior	PFN posterior-predictive used as initial CDF, followed by Gaussian-copula updates	Tabular uncertainty quantification
TabMGP	TabPFN used directly as the predictive rule within MGP	Supervised regression and classification on tabular data
MPNP	Neural-process predictive distribution used as the source of martingale-posterior uncertainty	Neural Processes
MPD	Guided diffusion bridges and bridge-based score estimates	Discretely observed diffusion processes
FMP	One-shot federated protocol using trainable data embeddings and centralized predictive sampling	Federated Bayesian neural networks

Copula-based predictive recursions were part of the original program. Given a predictive density $P_i$8 and CDF $P_i$9, a copula density $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$0 can be used to define

$F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$1

or its smoothed weighted variant, with the martingale property inherited from the copula construction (Fong et al., 2021).

The semiparametric moment martingale posterior uses the mixture predictive

$F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$2

with $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$3 chosen by method-of-moments matching. The predictive moments are martingales, the plug-in parameter and predictive CDF converge almost surely, and an energy-score-based leave-one-out procedure selects the mixture weight $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$4 (Yung et al., 24 Jul 2025).

The quantile martingale posterior tracks a quantile curve $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$5 and updates it by a martingale recursion driven by a Gaussian-copula kernel, with increasing rearrangement enforcing monotonicity. It yields an MCMC-free and parallelizable posterior computation scheme, and an asymptotic Gaussian-process approximation further accelerates sampling (Fong et al., 2024).

Two PFN-based variants are distinguished in the literature. One begins from a PFN posterior-predictive CDF $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$6 and then imposes martingale-preserving Gaussian-copula updates because naïvely iterating a PFN can violate the martingale property. The other, TabMGP, plugs TabPFN directly into the MGP framework: in supervised regression or classification, each $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$7, $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$8 is drawn by Bayesian bootstrap from the observed features, and TabPFN’s in-context API returns $F_N=\frac1N\sum_{i=1}^N\delta_{z_i},$9 (Nagler et al., 16 May 2025, Ng et al., 29 Oct 2025).

Neural-process and diffusion variants illustrate the range of admissible predictive rules. Martingale Posterior Neural Processes equip a Neural Process with a predictive distribution implicitly defined with neural networks and use the corresponding martingale posteriors as the source of uncertainty (Lee et al., 2023). For discretely observed diffusions, the MPD algorithm replaces repeated likelihood evaluation by bridge-based score updates, using guided auxiliary processes, Girsanov weights, and a two-phase warm-up plus generative continuation scheme (Yao et al., 30 Apr 2026).

In federated learning, federated martingale posterior sampling uses a one-shot protocol in which each client uploads a small set of trainable data embeddings and the server runs the predictive sampler centrally. This is presented as a federated instantiation of predictive Bayes for Bayesian neural networks (Zhang et al., 18 May 2026).

5. Empirical behavior across models and applications

Empirical evaluations emphasize calibration, coverage, and computational efficiency rather than posterior normalizing constants. In TabMGP, evaluation was conducted on 30 data-generating setups—11 synthetic and 19 real-world tabular datasets—with comparisons against Bayesian bootstrap, copula-based MGP, and classical Bayesian posterior via NUTS with Gaussian prior. The reported metrics were joint coverage versus trace of posterior covariance, marginal coverage, and Winkler score. For linear regression, TabMGP achieved near-nominal, often $\theta=\Theta(F)$0, coverage with the smallest to moderate set sizes; for logistic regression, it yielded good coverage, at least $\theta=\Theta(F)$1 in real data, with the smallest credible sets; and for marginal intervals it led in Winkler score on approximately 155 of 293 parameters. Convergence diagnostics based on trace plots of $\theta=\Theta(F)$2 confirmed $\theta=\Theta(F)$3 in practice (Ng et al., 29 Oct 2025).

The PFN-based uncertainty-quantification work reports a different empirical profile. It checks the martingale property directly, shows that naïvely iterating a PFN can produce empirical-CDF drift, and uses Gaussian-copula updates to enforce coherence. Reported experiments include convergence stabilization by $\theta=\Theta(F)$4–200 in Gamma examples, conditional regression posteriors for diffusion-process and subspace-inference data, conditional quantile coverage studies, bootstrap comparisons on UCI tabular datasets, unconditional $\theta=\Theta(F)$5 quantile posteriors for Gamma data, and an LLM proof-of-concept in which GPT-4o supplied approximate CDF prompts (Nagler et al., 16 May 2025).

For discretely observed diffusions, MPD is illustrated on Ornstein–Uhlenbeck and Lotka–Volterra examples. In the Ornstein–Uhlenbeck example, Phase 1 rapidly moved from $\theta=\Theta(F)$6 to near the true value $\theta=\Theta(F)$7 in 100 steps, and over 100 independent runs the final mean was approximately $\theta=\Theta(F)$8. In the Lotka–Volterra example, $\theta=\Theta(F)$9 converged quickly while $\theta(F_\infty)$0 converged more slowly but were still well estimated. Against MLPMMH on the Lotka–Volterra data, MPD reduced total wall time from approximately $\theta(F_\infty)$1 minutes to $\theta(F_\infty)$2 minutes, with cost per approximate independent draw falling from about $\theta(F_\infty)$3 seconds to $\theta(F_\infty)$4 seconds, roughly a $\theta(F_\infty)$5 speed-up (Yao et al., 30 Apr 2026).

Federated martingale posterior sampling was evaluated on MNIST, CIFAR-10, and CIFAR-100. Under iid partitions, FMP closely matched centralized MP and improved calibration over consensus-style baselines; under heterogeneous Dirichlet splits, it consistently achieved the highest accuracy and lowest expected calibration error among federated methods, especially at $\theta(F_\infty)$6 (Zhang et al., 18 May 2026).

In neural-process settings, Martingale Posterior Neural Processes were reported to outperform baselines on 1D regression, high-dimensional regression, image completion, and Bayesian optimization benchmarks. In deep neural networks, score-based martingale posteriors were competitive with NUTS on a small 20-parameter toy network but showed that preconditioner quality is decisive on modern convnets: block or full Fisher approximations were effective in small models, while diagonal Fisher variants over-dispersed severely on MNIST-scale networks (Lee et al., 2023, Zhumekenov et al., 14 Jun 2026).

6. Limitations, unresolved theory, and methodological distinctions

The predictive rule is the central design choice, and the utility of martingale posteriors depends on that choice. One recurrent theme is that the literature had offered few compelling examples of predictive rules before the use of foundation transformers and related black-box predictors. TabMGP explicitly argues that foundation transformers are well suited because their autoregressive generation mirrors forward simulation and their general-purpose design enables rich predictive modeling (Ng et al., 29 Oct 2025).

At the same time, theory and practice are not yet fully aligned. TabPFN fails exact martingale or almost c.i.d. conditions, yet TabMGP converges and attains nominal coverage. This suggests that existing sufficiency results are conservative and motivates development of weaker, more realistic convergence conditions for black-box predictive models (Ng et al., 29 Oct 2025).

A second limitation is structural rather than computational: one-step martingale coherence does not imply full multi-step predictive determination. In exchangeable Bernoulli models, the martingale posterior framework constrains only the first conditional moment of the terminal value unless additional structure is specified. Under any strictly proper scoring rule, the plug-in predictive is strictly dominated by the Bayes predictive whenever the posterior is non-degenerate (Polson et al., 28 Feb 2026).

Domain-specific implementations introduce further constraints. In TabMGP, continuous responses are binned, column-permutation invariance is enforced by ensemble averaging, and forward sampling requires repeated calls to TabPFN; an autoregressive training variant is identified as a possible remedy (Ng et al., 29 Oct 2025). In PFN-based martingale posteriors more generally, the martingale correction is external to the PFN because the base model need not satisfy coherent sequential updating on its own (Nagler et al., 16 May 2025).

Score-based and deep-learning variants carry their own sensitivities. The score-driven formulation depends on step sizes satisfying $\theta(F_\infty)$7 and $\theta(F_\infty)$8, while the deep-network exposition stresses that preconditioner quality is critical, that step-size tuning is delicate and architecture-dependent, and that extending score-based martingale posteriors to include explicit prior penalties remains open (Cui et al., 3 Jan 2025, Zhumekenov et al., 14 Jun 2026).

In diffusion problems, MPD requires simulating diffusion bridges in low to moderate dimension, and both step-size tuning and guide-process choice affect convergence. In federated settings, no formal privacy analysis is provided, embeddings may leak information, and meta-training requires a corpus of related tasks (Yao et al., 30 Apr 2026, Zhang et al., 18 May 2026).

Taken together, these results suggest a framework with a stable conceptual core—predictive specification, martingale convergence, and posterior sampling by forward imputation—but with validity and performance governed by the coherence, richness, and computational tractability of the predictive rule itself.