Sequential Bayesian Updating

• Sequential Bayesian updating is a method that uses recursive incorporation of new data to update the posterior distribution while retaining prior beliefs.
• The approach employs techniques like Monte Carlo and particle filtering to approximate complex posterior distributions and prevent particle depletion.
• Applications include streaming inference, online learning in deep neural networks, and hierarchical models, offering efficient computation for big data.

Sequential Bayesian updating is a fundamental methodology for incorporating new data into probabilistic models as it arrives, with the posterior distribution from each update serving as the prior for the next. This paradigm underpins streaming inference, big data partitioning, online learning in deep neural networks, population dynamics in hierarchical models, recursive estimation in state-space models, and decision-theoretic frameworks for sequential experimentation. Rigorous mathematical formulations, algorithmic strategies to avoid degeneracy, and practical diagnostics have been developed for both parametric and high-dimensional/nonparametric settings.

1. Formal Structure of Sequential Bayesian Updating

Let data arrive in batches $y_1, y_2, \ldots, y_T$, and let the model parameter vector be $\theta$. The core recursive rule is

$$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$

where $\pi_{t-1}(\theta)$ is the prior or "transient posterior" at step $t-1$, and $L(y_t\,|\,\theta)$ is the likelihood for the new batch $y_t$. In practice, neither $\pi_{t-1}$ nor $\pi_t$ typically admits a closed form, necessitating Monte Carlo, variational, or other approximate representations of the current belief state (Scharf, 3 Aug 2025). Posterior representations are carried forward either as a collection of samples (particles), analytic approximations, or variational parameterizations (Kochurov et al., 2018, Tomasetti et al., 2019, Scharf, 3 Aug 2025).

2. Monte Carlo and Particle-Based Algorithms: The SPP-RB Approach

In sequential settings, particle-based approximations are widely used. At each update, the current posterior is represented by $N$ particles $\theta$0. The smoothed prior–proposal recursive Bayes (SPP-RB) scheme introduces a kernel-smoothed mixture proposal

$\theta$1

with parameter $\theta$2, $\theta$3 determining the degree of shrinkage from a global Gaussian ($\theta$4) to pure kernel density estimation ($\theta$5). This approach ensures continuous support for the proposals and avoids "particle depletion," in which resampling collapses diversity (Scharf, 3 Aug 2025).

The SPP-RB method employs a Metropolis–Hastings (MH) within-Gibbs update:

1. For each particle, propose $\theta$6.
2. Compute MH acceptance ratio $\theta$7.
3. Accept or reject accordingly, (optionally) followed by resampling.

This procedure maintains moment preservation and low variance in weights (variance $\theta$8 as $\theta$9) compared to multinomial resampling, where weight variance can increase without bound (Scharf, 3 Aug 2025).

Simulation studies demonstrate that SPP-RB with moderate to low $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$0 achieves Kolmogorov–Smirnov distances to all-at-once posteriors of $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$1–$$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$2, whereas raw (non-smoothed) approaches degrade rapidly (KS $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$3) and lose particle uniqueness (Scharf, 3 Aug 2025).

3. Theoretical Guarantees, Computational Complexity, and Degeneracy Avoidance

Sequential updating is computationally attractive for large or streaming data, as per-iteration cost is $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$4 for each batch of size $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$5, much less than all-at-once $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$6. SPP-RB adds only an $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$7 per-particle cost for sampling mixture proposals, maintaining overall efficiency (Scharf, 3 Aug 2025).

Smoothing guarantees unique, diversified support for each proposal, and asymptotically, as $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$8 and $$\pi_t(\theta) = p(\theta\,|\,y_{1:t}) \propto L(y_t\,|\,\theta)\;\pi_{t-1}(\theta)$$9, $\pi_{t-1}(\theta)$0. Variance in importance weights is minimized, ensuring robustness against collapse to a single mode or loss of representational diversity that plagues naive approaches (Scharf, 3 Aug 2025).

High-dimensional settings motivate block-wise updates, where sub-vectors of $\pi_{t-1}(\theta)$1 are updated conditional on others, again with mixture proposals constructed from earlier-stage samples. Such blocked updating can use conditional mixtures weighted by the density of frozen coordinates (Scharf, 3 Aug 2025).

4. Diagnostic Tools and Practical Recommendations

Robustness and accuracy require diagnostics for particle degeneracy and posterior calibration. Repeated random data partitions followed by cross-comparison of resulting posteriors can detect insufficient particle size $\pi_{t-1}(\theta)$2 or miscalibrated shrinkage $\pi_{t-1}(\theta)$3. For high-dimensional or non-Gaussian/multimodal posteriors, adaptive schemes for bandwidth or shrinkage selection and hybridization with alternative MCMC kernels (e.g., slice-sampling) are recommended (Scharf, 3 Aug 2025).

Choosing $\pi_{t-1}(\theta)$4 (global Gaussian proposals) often leverages the Bernstein–von Mises theorem in moderate dimensions (asymptotic normality), while increasing $\pi_{t-1}(\theta)$5 enhances support for strong multimodality (Scharf, 3 Aug 2025).

5. Extensions, Limitations, and Model-Class Generality

Extensions include stage-adaptive bandwidth schemes, kernel choice generalization beyond the Gaussian, and combination with mini-batch sub-sampling for massive-data streaming (Scharf, 3 Aug 2025). The kernel smoothing can be seamlessly tuned from highly local to fully global, depending on the application’s geometry and information structure.

Limitations arise in fully nonparametric bandwidth selection in large $\pi_{t-1}(\theta)$6 (parameter dimension), where automatic calibration remains challenging. For extremely non-Gaussian, strongly multimodal targets with complex dependencies, further algorithmic sophistication may be necessary.

The SPP-RB architecture supports recursive inference in streaming, partitioned, or mini-batched datasets, hierarchical models, and in scenarios where full-data re-analysis is impractical.

6. Applications and Empirical Case Studies

Emphasizing the practical utility, SPP-RB is validated in simulations on logistic regression (with $\pi_{t-1}(\theta)$7 across $\pi_{t-1}(\theta)$8 batches) and high-dimensional hierarchical forest-classification models (covariate dimensionality $\pi_{t-1}(\theta)$9). In both, SPP-RB recapitulates the full-data posterior (KS distances $t-1$0) and retains the full particle ensemble even into late-stage updates, while naive methods collapse (Scharf, 3 Aug 2025).

In the forest-classification example, SPP-RB (with global Gaussian proposals) precisely tracks multivariate marginal contours and maintains particle count, outperforming both raw-PP-RB and univariate KDE-based proposals.

7. Summary Table: Key Properties of SPP-RB

Feature	SPP-RB	Raw PP-RB
Particle depletion	Avoided by smoothing, always $t-1$1 unique particles	Collapses to few unique points
Moment preservation	Yes, as $t-1$2, $t-1$3	Poor under repeated resampling
Proposal support	Full (continuous); exact as $t-1$4	Discrete, at old particles
Weight variance	$t-1$5 (vanishes for $t-1$6)	Constant or increasing
Extra computational cost	Minimal ($t-1$7 per sweep)	$t-1$8
Tuning parameter	Shrinkage $t-1$9, bandwidth $L(y_t\,|\,\theta)$0	None
Multimodal support	Adjustable via $L(y_t\,|\,\theta)$1	Collapses to dominant mode

SPP-RB offers a flexible, high-fidelity, and computationally efficient toolkit for streaming or partitioned Bayesian inference, with systematic mechanisms to prevent degeneration while maintaining statistical accuracy and moment fidelity at all stages (Scharf, 3 Aug 2025).