Bayesian Pooled Posterior Maximisation

• Bayesian pooled posterior maximisation is a modular approach that recombines independent posterior analyses into a pooled object for joint decisions or refined approximations.
• It is applied in settings such as hierarchical Bayes (MBA), likelihood-free inference via linear opinion pooling, multi-agent decision making with Blackwell-optimal aggregation, and moment matching in QEM.
• Empirical evaluations show that these methods achieve comparable or improved accuracy with significant computational speed-ups compared to full joint Bayesian inference.

Bayesian pooled posterior maximisation denotes a family of inferential constructions in which posterior information is generated separately and then recombined into a pooled posterior object, pooled decision rule, or pooled approximate posterior. In the cited literature, the expression is used in at least four technically distinct settings: a two-stage approximation for hierarchical Bayes based on combining independent source-specific posteriors; a linear-opinion-pool approach for likelihood-free inference (LFI); a Blackwell-theoretic upper-bound rule for multi-agent decision making based on the posterior under pooled private information; and a moment-matching procedure for approximate posteriors driven by massively parallel importance weighting (Dutta et al., 2016, Frazier et al., 2022, Zhang et al., 7 May 2026, Heap et al., 11 Mar 2025). A common feature is that pooling occurs at the posterior level rather than through a single monolithic likelihood evaluation, but the resulting object is not uniformly an exact Bayesian posterior in the original model.

1. Conceptual scope and unifying structure

In hierarchical modeling, pooling is performed by first sampling each source-specific posterior independently and then treating those posterior draws as observed data in a substitute hierarchical model with a scaled likelihood (Dutta et al., 2016). In LFI, pooling is performed by combining multiple approximate posteriors through a linear opinion pool, with the weight chosen to optimize the asymptotic frequentist risk of the pooled posterior mean rather than by equal weighting (Frazier et al., 2022). In multi-agent decision making, pooled posterior maximisation refers to the decision rule that predicts with the posterior $\Pr(y \mid x, d_{1:M})$ based on the agents’ full pooled private information, which is presented as the information-theoretic upper bound under Blackwell’s ordering (Zhang et al., 7 May 2026). In approximate-posterior learning, QEM pools information from a reweighted set of latent samples into posterior moment estimates and then updates the approximate posterior by moment matching rather than gradient ascent (Heap et al., 11 Mar 2025).

These uses are related but not identical. In some cases the pooled object is a posterior distribution over latent parameters; in others it is a posterior-derived point estimator or action rule. This suggests that “Bayesian pooled posterior maximisation” functions less as the name of one standardized algorithm than as a broader design pattern: independent or modular posterior analyses are retained, and a second operation aggregates them to recover a joint decision, a shared hyperparameter structure, or a refined approximation.

2. Two-stage hierarchical Bayes via independent posterior combination

The hierarchical-model formulation begins from

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$

with full posterior

$$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$

Direct inference requires joint sampling of all $\theta_j$ and $\phi$, typically by Gibbs or other MCMC, and becomes expensive when $J$ is large, each source model is complex, or the posterior is high-dimensional or mixes poorly (Dutta et al., 2016).

The proposed two-stage procedure, called meta-analysis of Bayesian analyses (MBA), first fits each source independently and obtains posterior samples

$$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$

from $\pi(\theta_j \mid \mathbf{x}_j)$. It then defines

$$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$

models $\psi_j$ hierarchically through $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$0, and treats the posterior samples $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$1 as if they were data generated from a distribution centered at $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$2.

The substitute likelihood is

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$3

The exponent $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$4 is critical: it prevents the substitute likelihood from becoming arbitrarily sharp as the number of posterior draws increases, and turns the product into a geometric average likelihood. The resulting substitute posterior is

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$5

The update equations mirror the Gibbs structure of the full hierarchical model: $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$6

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$7

The paper further justifies $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$8 heuristically by showing that it approximates the original source-posterior structure, with the central approximation that the substitute likelihood behaves like the original data likelihood times the prior on source-specific parameters.

Closed forms are available in conjugate settings. For multivariate normal source parameters with $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$9 and $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$0, the conditional posterior is $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$1 with

$$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$2

An inverse Wishart analogue is also given, and these closed forms make the second-stage pooling very fast.

The computational rationale is explicit. Source-wise inference is embarrassingly parallel; the source-specific subproblems are lower-dimensional and often converge faster; the second-stage reconstruction is usually simple and conjugate; and direct full-hierarchy bottlenecks are avoided. In simulation, for a normal-normal hierarchy with $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$3 and $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$4, MBA and full hierarchical Bayes had very similar MSEs and credible interval coverage, while MBA was about twice as fast on average, with FHM at $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$5s and MBA at $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$6s. In an inverse Wishart hierarchy with $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$7, FHM was around $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$8s and MBA around $$\pi(\phi,\theta_1,\ldots,\theta_J \mid \mathbf{x}_1,\ldots,\mathbf{x}_J).$$9s. In the retail cheese sales example with weekly sliced cheese sales in $\theta_j$0 retail stores, full hierarchical inference in Stan/HMC took about 1 hour, whereas the combined MBA analysis took about 5 seconds for $\theta_j$1 posterior draws after burn-in, roughly 300-fold faster, while producing posterior distributions for $\theta_j$2 close to those from full hierarchical Bayes (Dutta et al., 2016).

3. Risk-optimised posterior pooling in likelihood-free inference

In LFI, pooled posterior maximisation is introduced to avoid committing to a single summary statistic vector, discrepancy, or even a single LFI algorithm (Frazier et al., 2022). Standard LFI approximates a partial posterior $\theta_j$3 using summaries $\theta_j$4, but different summaries can yield different posterior centers, variances, and misspecification behavior, while high-dimensional summary vectors are costly and statistically unstable because of the curse of dimensionality.

For two summary sets $\theta_j$5 and $\theta_j$6, the pooled posterior is defined by the linear opinion pool

$\theta_j$7

Its posterior mean is

$\theta_j$8

The method is not standard Bayesian updating on a joint likelihood from $\theta_j$9; it is a mixture of posterior approximations. It is also not ordinary equal-weight posterior averaging, because the weight is selected to optimize asymptotic risk.

Because the plain linear pool can inflate posterior variance when component posteriors have different locations, a recentered pool is also defined. The recentered construction preserves the pooled posterior mean but yields

$\phi$0

with $\phi$1 in the positive-semidefinite sense. For uncertainty quantification, the paper recommends the recentered pool.

The optimization criterion is the asymptotic expected loss of the pooled posterior mean,

$\phi$2

which reduces to asymptotic mean squared error under quadratic loss. Under asymptotic normality of the summaries, injectivity and regularity of the summary mean map, a BvM-type result for each component posterior, and smooth loss, the paper derives an exact optimal weight $\phi$3 and proves

$\phi$4

A particularly important result concerns incompatibility: if one summary set is compatible and another is incompatible, then asymptotically the pooled posterior discards the incompatible one, with

$\phi$5

Because estimating the covariance term $\phi$6 is hard and noisy, two practical alternatives are recommended: $\phi$7 and

$\phi$8

These require only posterior draws from the component LFI posteriors. The paper remarks that estimating $\phi$9 can hurt finite-sample performance, a phenomenon described as akin to the forecast combination puzzle.

The method is explicitly designed as a wrapper around existing algorithms, including ABC, Bayesian synthetic likelihood (BSL), and discrepancy-based ABC methods such as Wasserstein ABC. Empirically, in the g-and-k model, the pooled posterior using $J$0 achieved total MSE $J$1, compared with $J$2 for $J$3, $J$4 for $J$5, and $J$6 for the expensive joint analysis. In the stochastic volatility model, pooled methods substantially beat both individual posteriors and the joint posterior, with total MSE $J$7 for $J$8 and $J$9 for $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$0, versus $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$1 for $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$2, $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$3 for $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$4, and $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$5 for the joint summary analysis. In M/G/1, pooling also improved over either BSL or Wasserstein ABC alone, with total MSE $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$6 for $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$7 versus $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$8 for $$\boldsymbol{\theta}_j^*=(\theta_{j1}^*,\ldots,\theta_{jL}^*)$$9 and $\pi(\theta_j \mid \mathbf{x}_j)$0 for $\pi(\theta_j \mid \mathbf{x}_j)$1 (Frazier et al., 2022).

4. Blackwell-optimal aggregation in multi-agent decision making

In the multi-agent setting, pooled posterior maximisation is formulated through Blackwell’s informativeness framework (Zhang et al., 7 May 2026). A decision problem is defined as $\pi(\theta_j \mid \mathbf{x}_j)$2, where $\pi(\theta_j \mid \mathbf{x}_j)$3 is the state space, $\pi(\theta_j \mid \mathbf{x}_j)$4 the action space, $\pi(\theta_j \mid \mathbf{x}_j)$5 the utility of action $\pi(\theta_j \mid \mathbf{x}_j)$6 in state $\pi(\theta_j \mid \mathbf{x}_j)$7, and $\pi(\theta_j \mid \mathbf{x}_j)$8 the prior on states. An information structure is $\pi(\theta_j \mid \mathbf{x}_j)$9 with signal space $$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$0 and conditional signal distribution

$$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$1

Given the prior, the posterior is

$$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$2

The expected value of an information structure under optimal decision making is

$$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$3

If $$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$4 is a garbling of $$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$5, written

$$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$6

then every Bayesian decision maker weakly prefers $$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$7 to $$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$8, so $$\psi_j \triangleq \mathbb{E}(\theta_j \mid \mathbf{x}_j),$$9.

The paper maps multi-agent QA and classification into this framework by identifying the state with the true label $\psi_j$0, the action with the prediction $\psi_j$1, and each agent’s private information with a signal $\psi_j$2. The joint private information is $\psi_j$3, with information structure $\psi_j$4. Any aggregation

$\psi_j$5

induces a garbling of the pooled private information, because $\psi_j$6 is a deterministic function of $\psi_j$7. Hence ensemble voting rules and multi-round debate are both less informative than access to the full pooled private information.

The resulting optimal decision rule is

$\psi_j$8

which the paper explicitly identifies as the maximum Bayesian pooled posterior. The upper-bound claim is decision-theoretic rather than purely empirical: because any aggregation of private information is a garbling, no voting or debate protocol can dominate access to the pooled source under Blackwell’s ordering.

For practical approximation, the paper assumes conditional independence of agents’ private information given label and input,

$\psi_j$9

and a uniform prior $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$00. Then

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$01

This yields the product-of-posteriors estimator

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$02

The practical algorithm, MA-PoP (Multi-Agent Pooled Posterior), estimates each agent’s posterior by Monte Carlo response sampling. For each agent $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$03, it samples $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$04 responses $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$05 and estimates answer-option probabilities through semantic similarity: $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$06 The method uses an NLI cross-encoder—specifically nli-deberta-v3-large—for semantic compatibility scoring, and a Deep Sets calibration head to ensure permutation equivariance with respect to option ordering.

On six QA benchmarks—MMLU Professional Medicine, MMLU Formal Logic, HellaSwag, CommonsenseQA, HH-RLHF, and MedMCQA—MA-PoP outperformed single-agent baselines, majority voting, self-consistency, log-linear opinion pooling, inverse surprising popularity, and several MAD variants. In the 5-agent heterogeneous benchmark table, MA-PoP achieved accuracy 0.8787 on MMLU Professional Medicine, 0.7367 on MMLU Formal Logic, 0.8433 on HellaSwag, 0.8800 on CSQA, 0.5900 on HH-RLHF, and 0.6467 on MedMCQA. Calibration improved markedly with the Deep Sets head; on MedMCQA, Falcon-7B ECE dropped from 0.2740 to 0.0871, and Falcon-34B ECE dropped from 0.1973 to 0.0305. The paper also reports that increasing Monte Carlo samples from $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$07 to $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$08 improves performance, after which gains saturate, and uses $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$09 as a practical balance (Zhang et al., 7 May 2026).

5. Moment-pooled approximate posterior fitting with QEM

QEM reframes approximate-posterior learning as an EM-like procedure in which the E-step computes posterior moments using massively parallel importance weighting (MPIW), and the M-step fits the approximate posterior by setting its moments equal to those estimates (Heap et al., 11 Mar 2025). The key object is the posterior

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$10

and the target of the E-step is $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$11 for suitable moment functions.

In the massively parallel setting, each latent variable $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$12 has $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$13 samples, and the method considers all $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$14 combinations. The MPIW estimator produces pooled posterior moment estimates from this reweighted sample set. QEM then replaces gradient ascent with direct moment matching. For an exponential-family approximate posterior

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$15

the mean parameters are

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$16

Instead of updating natural parameters by a gradient step, QEM updates mean parameters through

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$17

equivalently

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$18

The paper emphasizes that QEM is not standard EM: standard EM updates model parameters in the prior or likelihood, whereas QEM updates the parameters of the approximate posterior $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$19.

The supported approximate posterior families explicitly include Gaussian, Gamma, Beta, Dirichlet, Binomial, Multinomial, Categorical, and combinations thereof. For a scalar Gaussian, the update is given through the first two moments: $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$20 For Beta and Gamma families, the fitting uses moment or log-moment statistics; for Dirichlet, expected log-components; and for discrete exponential families, posterior category or success probabilities.

QEM is presented as a form of Bayesian pooled posterior maximisation because it pools information from reweighted latent samples into posterior moment estimates and updates the approximate posterior by moment matching rather than ELBO gradient optimization. The method is designed to exploit conditional-independence structure, with appendix-level memory reduction from

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$21

An asymptotic theorem states that if $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$22 is unbiased with finite variance and

$$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$23

then asymptotically $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$24 is unbiased and its variance goes to zero. The appendix also states that as $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$25, massively parallel importance weighting gives the correct posterior moments under mild support conditions, and therefore QEM converges in a single step in that limit.

Empirical evaluation covers Bus Breakdown, MovieLens100K, Bird Occupancy, Radon, and Covid. Baselines are massively parallel VI and massively parallel RWS, with Adam learning rates selected from $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$26. Across 250 inference iterations and 5 random seeds, QEM outperformed MP VI on ELBO and predictive log-likelihood in every model, learned faster than MP RWS in all settings except Radon and Occupancy on predictive log-likelihood, had lower standard error across repeated runs, and was typically faster per iteration. Reported times for 250 iterations include 11.7s for QEM versus 18.6s for MP RWS and 19.7s for MP VI on Bus Breakdown, and 19.8s for QEM versus 49.6s for MP RWS and 62.7s for MP VI on MovieLens. A major theoretical and empirical claim is reparameterization invariance: QEM remains essentially unchanged under latent-variable rescalings that slow down or destabilize VI and RWS (Heap et al., 11 Mar 2025).

6. Exactness, assumptions, and recurrent limitations

A persistent misconception is that pooled posterior maximisation always reproduces exact Bayesian conditioning on all original data. The cited literature does not support that interpretation. In MBA for hierarchical models, the substitute hierarchical model is explicitly an approximation based on a parametric density $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$27 fitted to source posterior samples, and the paper identifies nonparametric representations of source posteriors as future work (Dutta et al., 2016). In LFI, the linear opinion pool is explicitly not conditioning on a joint likelihood from $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$28, but a mixture of posterior approximations chosen to optimize asymptotic risk (Frazier et al., 2022). In the multi-agent setting, the Blackwell-optimal pooled posterior is conceptual, while the product-of-posteriors implementation is exact only under conditional independence of agents’ private information and a uniform prior; dependence induces overconfidence, although the reported empirical degradation is gradual rather than catastrophic (Zhang et al., 7 May 2026). In QEM, the procedure still approximates the posterior through importance weighting and therefore inherits the usual support-mismatch and weight-degeneracy limitations of importance-sampling methods (Heap et al., 11 Mar 2025).

Another recurrent issue concerns the distinction between pooling for point estimation and pooling for uncertainty quantification. The LFI paper centers its guarantees on the asymptotic frequentist risk of the pooled posterior mean, not on a universal optimality statement for the full pooled distribution. This is why the recentered pool is preferred when posterior variance inflation matters (Frazier et al., 2022). The multi-agent paper likewise separates the Blackwell-optimal decision rule from the practical posterior estimator and emphasizes calibration as a necessary complement to pooling (Zhang et al., 7 May 2026).

The approximation quality is strongly family-dependent. MBA works best when the source-specific posterior distributions are reasonably well captured by the chosen parametric family $$X_{ji} \sim F_{X|\theta}(\cdot \mid \theta_j), \qquad \theta_j \sim F_{\theta|\phi}(\cdot \mid \phi), \qquad \phi \sim F_\phi,$$29 and when the posterior is not highly irregular (Dutta et al., 2016). The LFI theory is cleanest when component summaries are compatible and each component posterior satisfies a BvM-type asymptotic regime; if all summaries are incompatible, the authors recommend robust LFI methods instead (Frazier et al., 2022). QEM is straightforward only for exponential-family approximations whose parameters are recoverable from moments, and raw moment updates can be unstable when a single importance sample dominates, which motivates EMA smoothing (Heap et al., 11 Mar 2025).

Taken together, these works define Bayesian pooled posterior maximisation not as a single doctrine of posterior combination, but as a technically heterogeneous class of strategies for modular Bayesian aggregation. In each case, the attraction is the same: separate analyses can be run in parallel or on lower-dimensional objects, and a subsequent pooling step attempts to recover statistical efficiency, shrinkage, informativeness, or approximate joint inference without constructing one large coupled inference procedure from the outset.